※ 以前のバージョンの説明はAbout this versionから確認してください。
※ Please refer to "About this version" for details on the previous version.
noobaiの系列、illustriousの系列を問わず新しいモデルのいくつかを軸に一からマージをし直しました。そのためV4以前とは全く違う挙動をします。
新しいモデルを取り入れたため縦長ならば少し大きいサイズの出力でも問題がありません。(横長は胴が長くなりやすいです)
I have merged several new models, regardless of whether they belong to the Noobai or Illustrious series. As a result, the behavior is completely different from V4 and earlier versions.
Due to the incorporation of new models, vertically long images can now be generated with relative stability, even at slightly larger sizes. (However, in horizontally long images, the torso tends to appear elongated.)
目が崩れやすいので余裕があればADetailerの顔への適用をオススメします。
また確認できる問題点として暗闇の描写がやや苦手です。
Since the eyes can easily become distorted, if possible, it is recommended to apply ADetailer to the face.
One noticeable issue is that the model struggles somewhat with depicting darkness.
推奨する品質プロンプト/Recommended quality prompt
masterpiece, best quality, 推奨するネガティブプロンプト/Recommended negative prompt
worst quality, low quality, realistic, mutated hands, light particles, jpeg artifacts, lowres, signature, twitter username, company name, logo, sidelighting, heart-shaped pupils, 推奨するプロンプトをまず使用し、そこから調整をしてください。
また、特定のサンプラーや特定のスケジュール、CFGスケールが7を超えるとよくないためサンプルで使用した出力設定も参考にしてください。
例としてDPM++ 2M SDE HeunではNormalやUniformのスケジュールタイプを使用すると強烈にノイズが乗ります。
First use the recommended prompt, and then adjust it from there.
Additionally, certain samplers, specific schedules, and CFG scales above 7 do not work well, so please refer to the output settings used in the samples.
For example, with DPM++ 2M SDE Heun, using the Normal or Uniform schedule types results in excessive noise.
VAE内蔵済みです。
The VAE is already integrated.
サンプルで使用した出力設定/Output settings used for the samples
Sampling method:28
Output size:1040*1520
Schedule type:DPM++ 2M SDE Heun
Sampling steps:SGM Uniform
CFG Scale:5 ※7以上は推奨しません/A CFG scale above 7 is not recommended
Random number generator source:GPU
Emphasis mode:no norm
Eta noise seed delta:31337
CLIP skip SDXL:2
サンプルを出力した環境/Environment used to output samples
reForge
RTX3060
マージに使用したモデル/Models used for merging
Magic_ILL
waiNSFWIllustrious_v120
hikariNoobVPred_120
illustriousXL10_v10に手を改善するloraをマージしたモデル/A model merged with a LoRA that improves hand depiction, based on IllustriousXL10_v10
noobai0.75をベースに画風の学習をした自作のモデル/A custom model trained on art style, based on Noobai 0.75
songmix_v201
マージに使用したモデル及び、学習のベースとなったモデルの制作者達に感謝いたします。
I would like to express my gratitude to the creators of the models used for merging and the models that served as the base for training.
merge recipe
Magic_ILL x (1-alpha) + waiNSFWIllustrious_v120 x alpha = A
1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1A x (1-alpha) + hikariNoobVPred_120 x alpha = B
0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0B x (1-alpha) + illustriousXL1.0の手改善モデル x alpha = C
0,0,0,0.4,0,0.4,0,0.4,0,0,0.4,0,0,0,0,0,0.4,0.4,0,0C x (1-alpha) + noobai0.75をベースに画風の学習をした自作のモデル x alpha = D
0,0,0.6,0,0,0.6,0,0,0.6,0.6,0,0.6,0.6,0,0,0.6,0,0,0.6,0D x (1-alpha) + songmix_v201 x alpha = E
0,0,0,0,0,0,0,0,0.2,0.4,0.4,0.6,0.6,0.2,0,0,0,0,0,0E.fp16 x (1-alpha) + C x alpha = LibrasIllustriousXL_V5
0,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4Description
140枚で特化した画風チューニングをしたIllustriousXL0.1にマージを施し調整をしたモデルです。
(後にチューニングするのではなく前でチューニングをしたことになります。)
This is a model that was merged with IllustriousXL0.1, which underwent specialized style tuning with 140 images.
(the tuning was done before the merge, not after.)
マージレシピ/merge recipe
Illustrious tuning model x (1-alpha) + obsessionIllustrious_v10 x alpha [1,0,0,0,0,0,0,0,1,0,0,1,1,0,1,0,0,0,0,0] = AA x (1-alpha) + XL_e4P_Mix x alpha [0,0,0,0.3,0.6,0,0,0.3,0,0.3,0,0,0,0,0,0.3,0,0.3,0.3,0] = LibrasIllustriousXLIllustriousXL0.1及びobsessionIllustrious_v10のモデルの制作者様達に感謝いたします。
We would like to express our gratitude to the creators of the IllustriousXL0.1 and obsessionIllustrious_v10 models.
※特化したチューニングをしたためloraの効きが非常に悪い可能性があります。
※Due to specialized tuning, the LoRA effect may be significantly weak.
VAE内蔵済みです。
The VAE is already integrated.
目が崩れやすいのでADetailerの使用を推奨します。(サンプルの画像にも使用しています。)
CFG Scaleはあまり低くしすぎないようにしてください。低いと斑な画像になりやすいです。
The eyes tend to break easily, so we recommend using ADetailer. (This was also used in the sample images.)
Please avoid setting the CFG Scale too low. If it's too low, the image is more likely to become uneven.
サンプルで使用した品質プロンプト/Quality prompt used for the samples.
masterpiece, best quality, good quality, サンプルで使用したネガティブプロンプト/Negative prompt used for the samples.
worst quality, bad quality, average quality, traditional media, sketch, multiple views, bad anatomy, error, bad hands, blurry, text, jpeg artifacts, logo, artist name, watermark, chromatic aberration, signature, username, scan, abstract, 追記:bad handsの値は1.1にした方が指の安定性が上がる気がします。また、 jaggy lines, unclear, を追加すると滲みを軽減できます。
PS:I feel that setting the 'bad hands' value to 1.1 improves finger stability. You can reduce blurring by adding "jaggy lines, " and "unclear, ".
※まずはこのプロンプトの形から使用して調整してください。
※Please start by using this prompt format and make adjustments from there.
サンプルで使用した出力設定/Output settings used for the samples
Sampling method:30
Schedule type:Euler a
Sampling steps:Simple
CFG Scale:8
Random number generator source:NV
Emphasis mode:no norm
Eta noise seed delta:31377
サンプルを出力した環境/Environment used to output samples
reForge
RTX3060